农产品近红外光谱分析中的多模型共识方法研究

Using Consensus Modeling to Analyze Near Infrared Spectroscopic Data of Agricultural Products

基于多模型共识的基本思路结合近红外光谱,建立了多模型共识偏最小二乘回归方法(cPLS),从训练集随机取样建立一系列偏最小二乘回归模型(PLS),选取其中性能较好的部分模型作为成员模型,并用这些成员模型预测未知样品。将cPLS用于玉米中湿度、淀粉、蛋白质及油分含量的近红外光谱定量预测。结果 PLS对独立测试集中4种组分进行50次重复预测的平均预测误差均方根分别为0.020 7、0.268 6、0.122 0和0.070 6,预测误差均方根的标准偏差分别为4.753 0×10-3、0.054 8、0.023 0和0.014 9;而cPLS重复50次预测的平均预测误差均方根分别为0.016 0、0.167 8、0.116 6和0.044 1,预测误差均方根的标准偏差分别为2.735 0×10-4、0.002 5、0.003 0和7.683 0×10-4。可见,cPLS所建立的模型更加稳健可靠,预测的准确性也明显提高。

Based on consensus modeling combined with near infrared spectra（NIRS）, the consensus partial least squares（cPLS） regression method was established. A series of PLS regression models were built on training subsets constructed by random sampling from the training set. The models with high performance were selected as member models and used for pre- diction. The cPLS was used for modeling on NIRS data with moisture, starch, protein and oil of corn samples. The method was compared with the single-model PLS regression. Results showed that the single-model PLS regression obtained 0.020 7, 0.268 6, 0.122 0 and 0.070 6 of mean values of Root Mean Square Error of Prediction（RMSEP） on 50 repeat prediction for the four components on the independent test set. The standard deviation of the RMSEPs were 4.753 0×10-3, 0.054 8, 0.023 0 and 0.014 9, respectively. cPLS obtained 0.016 0, 0.167 8, 0.116 6 and 0.044 1 of mean RMSEP and 2.735 0×10-4, 0.002 5, 0.003 0 and 7.683 0×10-4of corresponding standard deviations. The results indicated that the models built by cPLS were more steady and reliable. The prediction results were more accurate than that of the single-model PLS regression.